Problem

Currently, when one defines a struct with an atomic, one has to use it as a read_write.

struct MiscData {
    num_fine_shapes: atomic<i32>,
    ...
};
@group(0) @binding(0) var<storage, read_write> buffer : MiscData;

If one tries to use it as a read only, which could make sense when passing the storage buffer to subsequent shaders, then one gets an error, due to how WGSL validation works at the moment.

// This is currently not possible, since it leads to an error
@group(0) @binding(0) var<storage, read> buffer : MiscData;

Error while parsing WGSL: :26:42 error: atomic variables in <storage> address space must have read_write access mode
@group(0) @binding(0) var<storage, read> buffer : MiscData;
                                         ^^^^^^^^^^^^^

:16:1 note: atomic sub-type of 'MiscData' is declared here
num_fine_shapes: atomic<i32>,
^^^^^

Current workaround

This means, if one wants to run multiple compute shaders that only read from an atomic variable in parallel, then one has to resort to creating an extra struct, with exactly the same layout and alignment, but a non-atomic variable.

struct MiscData {
    num_fine_shapes: i32, // I believe the spec says that this is valid
    ...
};
@group(0) @binding(0) var<storage, read> buffer : MiscData;

That is inconvenient, and could easily lead to either bugs (developer defined the secondary struct incorrectly) or performance left on the table (developer just used read_write).

Request

You could open an issue for that as a feature request. The validation on atomic built-in functions could be changed to have the access restrictions instead of the type. Only atomicLoad would be allowed in the read-only case. Hopefully, drivers would just optimize them to regular loads.
quote from alan-baker

Would it be possible to change the validation on atomic built-in functions to allow for atomicLoad in the read-only case?